Video generation has long faced a dilemma: the longer the generation time, the higher the compute cost of the full attention mechanism. Past optimizations have always traded quality for speed — but Sparse Forcing (arXiv:2604.21221) proposed by Meta and UCSB researchers proves sparse attention can boost both quality and speed.
The core insight: the attention of autoregressive diffusion models only concentrates on a small number of key visual blocks, forming implicit spatio-temporal memory. Based on this, the team designed PBSA (Persistent Block-Sparse Attention) — dynamically learning compression, retention, and update of persistent blocks, limiting computation to local windows.
The experimental data is counter-intuitive: for 5-second video, VBench +0.26, decoding speedup 1.11-1.17×, KV Cache peak memory down 42%. For longer 20-second video, +0.68 VBench, 1.22× speedup; for 1-minute video, +2.74 VBench, 1.27× speedup. The longer the time, the bigger the gain.
Why does quality actually improve? Forcing the model to learn which information is worth keeping is essentially structured regularization — reducing noise propagation and making content more coherent. PBSA's GPU kernel implementation also makes sparse computation truly usable.
For the industry: when video generation moves to minute-level scale, rather than piling on compute, let the model learn to be lazy — focus only on the truly important visual blocks. This is also a new direction for multimodal large-model long-context optimization.